摘要
为提高复杂交通场景下的目标检测效率,提出了一种基于改进YOLOv8模型的交通场景目标检测方法。将YOLOv8的主干网络替换为密集连接网络(DenseNet),以增强特征复用能力和梯度流动性;在SPPF模块前引入高效金字塔压缩注意力模块(PSA Module),实现多尺度特征聚合与通道权重自适应调整,从而强化上下文建模和显著区域表达能力;将损失函数替换为辅助边框损失函数(Unified-IoU),提升边界框拟合精度。结果表明,改进的YOLOv8模型在KITTI公共数据集上的精确率(Precision)达到88.2%,较YOLOv8n模型精确率提升4.5%,较YOLOv10n模型精确率提升2.1%;同时推理速度达98 fps,召回率为87.3%。该模型在车辆、行人等目标的检测精度上均有所提高,验证了其在交通场景目标检测中的优越性能和应用潜力。
To improve object detection efficiency in complex traffic scenarios,a traffic scene object detection method based on the improved YOLOv8 model is proposed.By replacing the YOLOv8 backbone with a densely connected network(DenseNet),feature reuse and gradient fluidity are enhanced.An efficient pyramid compressed attention module(PSA)is introduced before the SPPF module to achieve multi-scale feature aggregation and adaptive adjustment of channel weights,strengthening context modeling and salient region representation.The loss function is replaced with the Unified IoU loss function,which improves bounding box fitting precision.Experimental results show that the improved YOLOv8 model achieves a precision of 88.2%on the KITTI public dataset,a 4.5%improvement over the YOLOv8n model and a 2.1%improvement over the YOLOv10n model.The inference speed is 98 frames per second(fps),and the recall is 87.3%.The improved detection accuracy for objects such as vehicles and pedestrians demonstrates this paper's model's superior performance and potential for application in traffic scene object detection.
作者
钱睿珂
胡海霞
QIAN Rui-ke;HU Hai-xia(School of Mechatronics Engineering,Anhui University of Science and Technology,Huainan 232000,Anhui)
出处
《商洛学院学报》
2025年第6期43-54,共12页
Journal of Shangluo University
基金
安徽省淮南市科技计划项目(2023A3113)。
